So I have been reading some papers that came out recently, talking about forecasting markets and introducing new methodologies. I find huge mistakes that many novices and even professionals do in these models. Especially if you look on Youtube, there are plenty of amateurs there doing forecasting in spreadsheets or with R or other software, but they always make some kinds of mistakes that make their research completely bogus.

Yes, if you do only 1 mistake, then your entire model is rubbish. Historical data has to be handled carefully, because it's past data, so you have to simulate as if you were in the present, and the data next to it you know, but you have to pretend like you don't, since you are trying to simulate a model on something you don't know until the future comes. In a real-time model you don't know the future, so in the past model you should pretend that you don't know it either. So you can't leak into the future that is one mistake that many people do.

Don't Leak Into the Future

So on many Youtube videos they do things like, they calculate some statistic on the main data, like say the average price. And then they introduce weights according to that observation on a sample data, let me give you an example:

So in this example we have the BTC/USD price, we calculate the differences between the days, then we take the average of the differences, and then we add the average difference to each datapoint, to forecast the next datapoint from it.

Sounds easy right? Except that this forecast is bullshit. We have leaked into the future, by using the average of the differences. When we are at datapoint 1, we are not supposed to know anything about datapoint 2 and anything that comes in the future. Nothing!

At datapoint Aug 3, we are not supposed to know anything about anything that comes after that, because if that had been the present, we would not know anything about Aug 4 either. So we have to work as if we don't know anything about the future. So any tool that we use, can only be used on the data leading up to that point, but it can't incorporate information we don't know at that point.

This also applies to regressors as well. Suppose you use the BTC/USD data with some trade volume regressor, and the main BTC/USD data is aligned to 00:00 UTC midnight, but the regressor is aligned to 00:10 UTC midnight (I have literally seen sliding time like this in data).

And then you apply the regressor with the price, not knowing that the regressor leaks in the future by 10 minutes, which you are not supposed to know at 00:00 UTC time.

So aligning the time is also very important, it's a pain in the ass most of the time, since most of the data is really aligned like that. Some data, are represented in their own timezone, not even UTC. Now that is a real pain in the ass to align time then, since even something like a leap year or leap hour can mess up our analysis.

So I can't take any data seriously which isn't in UTC midnight format, if it's daily data. If hourly then also it should be aligned to the hour at 0 minutes.

So it's very important to only use past datapoints in your model, and also have those datapoints aligned by date and time correctly. Plus some models leak into the future by default, like the HP filter, so you can't use those incorporated into the model. The parameters has to be based on past data only.

So you can't introduce future information of any kind in your past data, that you didn't know at that moment in time.

Injective Transformation

The other mistake people do is by using bogus transformation functions. You can do anything you want with your data, if you follow the previous rule. Anything except introducing new information.

So transforming the data is sometimes useful to remove noise, or big spikes that distort it but have no meaning.

However the transformation function must be injective, possibly bijective, but at least injective.

This is to avoid bogus values at reconstruction. A non-injective surjective function can have 2 original values for 1 transformed value:

A good example of this is the Cosine function:

So you have the price in column A and you transform it through a cos(x) function in column C. You do your analysis on that. Now you want to transform it back to reconstruct the forecast, and you realize that C2,C20,C22 values can give different values. So which one is it? It can be any of them.

So this just introduces randomness in your model, which you don't want. So you can't use functions that are non-injective, probably all trigonometric functions are useless in forecasting. Not just that but any cyclical functions are probably useless too.

Good Benchmarking and Error Quantification

Another mistake is that people compare apples to oranges. So you have a model and you compare it's results to another model. I think that is wrong. Simply because different models use different lags or sizes of the data.

So I just introduced a new comcept, by comparing a model to it's random equivalent. The hypothesis is that the price is random, which is described by a white noise model ARIMA(0,0,0), basically taking the previous price as a forecast for the next price.

Then the error terms have to be un-biasedly described where I find RMS and things like that ridiculously biased, so I use LN ratio based errors.

Then just compare the LN Error of the model to the LN Error of it's "Random Walk" hypothesis. If the error is smaller then, the model is predictive, and gives us a positive edge. If it's equal then it has 0 edge. If bigger, then it's worse than randomness, a negative edge.

• https://steemit.com/trading/@profitgenerator/more-playing-with-btc-usd-data-quantitatively
• https://steemit.com/trading/@profitgenerator/the-real-structure-of-the-market
• https://steemit.com/mathematics/@profitgenerator/variability-in-a-heteroskedastic-market

So this is to avoid using things like Theil's U which is popular, but I think it's biased too.

"Technical Analysis"

So I have a distaste for TA, it's complete nonsense, mysticism based price analysis. Not science. But really apart from being non-factual and visual based, it can't be quantified, and the worst part is that TA analysts violate all these rules unconsciously.

For example the first rule is not just violated by visual TA analysts but also by science oriented quantitative analysts, but it's more hilarious seeing both people make those mistakes, it's more fun watching TA people drawing lines on the price showing some future price movement on the chart, that they have supposedly guessed, but they wouldn't have known about it if it wouldn’t have already happened.

They can draw nice lines on historical price, but they can’t predict shit about actual future data, and they always lose money in trading.

It’s interesting to see when brokers release their customer’s win/loss statistics, it shows that about 90% of traders consistently lose money, while only 10% of them make money and a few of them make really good money.

Almost like in a casino, except that a casino is provably negative expectancy, the house edge makes it probably negative.

A market is not negative expectancy, except for the tiny exchange fees like 0.1-0.2%, which normally should be negligible. But it’s mostly because these people don’t know how to forecast the market.

Prices can be forecasted, maybe even profitably, it’s just that they don’t know how, and they commit these mistakes. So trading is not necessarily gambling, only if they don’t know what they are doing.

---

Sources:
https://pixabay.com

---

Upvote, ReSteem &